•Minimum temperature in meteorological shelter that best describes frost is 2.5 °C.•Condition that the dew point temperature is less than 0 °C reduces false alarms.•Tmin threshold of 3 °C and Td of 0 °C gave the best results for frost probability.•These method could be coupled with climate models for frost prediction in future. Due to the earlier start of phenological cycles among fruit trees, frost represents one of the most notable hazards for agriculture. There is no unique method for forecasting frost, and different methods for describing frost under present and future climate conditions can be found in the literature. Often these methods are applied in a certain area without prior control. Five such frost detection methods were assessed in Croatia. In addition, five new frost estimation methods that rely on measurements of the daily minimum temperature (Tmin) and dew point temperature (Td), calculated using Tmin, relative humidity (RH), and the Clausius Clapeyron equation, as well as machine learning, were introduced in this research and compared to other methods. Overall, the frost prediction results showed that the minimum temperature measured at the meteorological shelter that best describes frost formation is 2.5 °C. Additionally, the condition whereby the dew point temperature is lower than 0 °C results in a reduction in the proportion of false alarms. Methods that introduce additional variables outperform those that rely solely on the temperature. The method in which days are classified as exhibiting frost using a Tmin threshold of 3 °C and Td threshold of 0 °C (ased on Tmin and daily mean RH) could capture the most frost days with the smallest error. This method is the most suited for continental areas with a high probability of detection (POD > 0.9) and a probability of false detection (POFD < 0.3) which conforms with the history of frost occurrence in this type of climate zone. These findings were corroborated by signal detection theory analysis, yielding high values of the accuracy index and beta values below 1, indicating a bias toward estimating frost events (with high hit rate values and high false alarm values). This method could be used to identify geographic areas most susceptible to frost formation and, if coupled with a climate model, enable the study of the frost vulnerability due to climate change.